[Abhishek-Mohan]

Which issue does this PR close?

Part of #6619. The tolerance-comparison documentation remains separate, so this does not close the issue by itself.

Rationale for this change

A SQL fixture can match Spark's results without exercising the expression its author intended to test. Disabling ConstantFolding does not stop NULL propagation, evaluation over inline VALUES, or Comet's planning-time literal casts. The guide also incorrectly showed binary bit_length using the codegen dispatcher.

What changes are included in this PR?

Update docs/source/contributor-guide/sql-file-tests.md to:

• Explain the remaining traps: NULL literals, literal casts, inline VALUES, small-insert file layout, regex escaping, and session time zones.
• Show a Parquet-backed string-to-int cast example with valid, invalid, and NULL inputs, and add the COALESCE(1) insert idiom to the setup steps without promising batch size or row order.
• Correct the ascii(NULL) example and the all-literal execution claim; replace the binary bit_length dispatch example with explicitly configured lower.
• Explain that operator coverage does not establish expression execution, including why expect_native can still report a planning-time literal cast as native.

How are these changes tested?

Documentation only. Checked the examples and claims against the SQL-file parser, runner, expression serdes, coverage tagging, existing fixtures, and Spark optimizer/parser sources.

• git diff --check passes.
• Structural checks pass for all six local heading links and balanced code fences.
• SQL runtime tests and the Sphinx build were not run.
• Prettier was not run because Node, npx, and Prettier are unavailable in the environment.

Council review

Rounds: 1

Delegate	Round 1
Claude Opus 5.5	9/10
Gemini 4 Argon	10/10
GPT-6 Astra	10/10

All three reviewers were satisfied with the scoped change and found no required fixes. Reviews followed the tstack council workflow at high reasoning effort.

AI assistance: implemented with Claude Opus 5.5 and coordinated with GPT-6 Astra through OpenCode in T3 Code.